Download Chapter 6 - Expert Systems and knowledge

University of Pretoria etd – De Kock, E (2003)
Chapter 6 - Expert Systems and
knowledge acquisition
An expert system’s major objective is to provide expert advice and knowledge in specialised situations
(Turban 1995). ES is a sub-discipline of AI (Turban et al 2001). For an ES to reason, provide
explanations and give advice, it needs to process and store knowledge. Knowledge in the form of facts
and rules are kept in a knowledge base.
6.1
Expert System definitions
ES definitions by Turban et al (2001) and Olson & Courtney (1992) were presented in Paragraph 2.3.1
(p19). Raggad & Gargano (1999) state: “ES emerged, as a branch of AI, from the effort of AI
researchers to develop computer programs that could reason as humans”. Goodall (1985) distinguishes
two approaches when defining an expert system:
•
The human/AI oriented approach, and
•
The technology oriented approach
Goodall’s (1985) AI approach defines an expert system as a computer system that performs functions
similar to those normally performed by a human expert. He extends this definition to include the
nature of an expert system, in other words, he includes how ES behaves, in the technology approach:
“An expert system is a computer system that uses a representation of human
expertise in a specialist domain in order to perform functions similar to those
normally performed by a human expert in that domain”.
In formulating the above definition, Goodall (1985) focused on the implementation of two types of ES:
•
ES performing at a suitable ‘high’ expert level, and
•
ES that performs tasks which no human expert has, using perhaps a very complex set of rules
to control a chemical process at a speed and in circumstances (e.g. temperature, time, location)
which no human could possibly match. Such a system will most surely outperform the best
human expert.
The value of ES as a sub-discipline of AI (Turban et al 2001), can be appreciated more when
comparing AI to natural intelligence and conventional programs. Kaplan (1984) as referred to by
Turban et al (2001), states several commercial advantages of AI over natural intelligence:
•
AI is more permanent. Employees take knowledge with them when they leave their place of
employment, or worse:
an employee may even forget his knowledge.
As long as the
computer systems and programs remain unchanged, the knowledge remains the same.
107
University of Pretoria etd – De Kock, E (2003)
•
AI offers ease of duplication and dissemination. Transferring knowledge between persons
may be a lengthy process. Some expertise is never transferred. Knowledge embodied in a
computer system can be easily moved to another computer.
•
AI can be less expensive than natural intelligence. Buying computer services can cost less
than having a human performing the same task.
•
AI is a computer technology and thus consistent and thorough. Natural intelligence is erratic
because people are erratic and perform inconsistently. , and
•
AI can be documented. Decisions made by a computer can easily be documented by tracing
the activities of the system. Natural intelligence is difficult to document. A conclusion
reached may be difficult to be recreated or recalled later. Some of the assumptions or even the
reasoning process may be forgotten.
Natural intelligence has several advantages over AI:
•
Natural intelligence is creative:
In an AI system, knowledge has to be carefully
constructed, while knowledge is inherent to all human beings
•
Natural intelligence uses sensory experienced directly: Users using AI, use sensory
experiences indirectly
•
Natural intelligence enables people to recognise relationships between artefacts, sense
qualities and spot patterns that explain how various items interrelate, and
•
A wide context of experiences is used when natural intelligence solves an individual
problem, while AI typically gains its power by having a very narrow focus
ES differs from conventional programs too. Raggad & Gargano (1999) and Turban et al (2001)
summarise the characteristics of conventional programs as follows:
•
Conventional Computer Based Information Systems (CBIS) processes data and information
as a direct conceptual resource
•
Conventional programs are based on algorithms or mathematical formulas and sequential
procedures that lead to a solution using data to solve problems
•
The objective of CBIS is to create and efficient solution by designing an efficient algorithm
•
CBIS are characterised by repetitive processing
•
The processing system or control mechanism of the system and the knowledge of the subject
are intertwined as part of the sequence of instructions
•
Knowledge of the subject is built into the processing mechanism or control of the system. The
subject knowledge or business rules form part of the sequence of instructions. All business
rules are part of the program code. The knowledge is stored as control embedded in the
program instructions. The control is the sequence in which the transactions are performed.
The programmer fixes these steps in advance. It is called the algorithm of the program.
•
To change the business rules or domain knowledge a programmer needs to edit the program
and insert or change the statements that represent the information in the correct place
108
University of Pretoria etd – De Kock, E (2003)
•
The program or system used does not explain which rules were used during processing of data
to produce the output
•
The program is specialised to deal with specific information. A complete system of programs
and input are necessary to achieve something useful.
•
One missing essential step causes the program(s) to malfunction. The data that are processed
should be complete and certain to achieve an appropriate result or valid output.
•
Error detection of the domain knowledge is performed when executing the program and
noticing faulty output
•
CBIS frequently uses algorithms in its search approach ,and
•
CBIS has no reasoning capability except for a fixed process in the algorithm of the program,
and
•
Updating and maintenance of data and information are usually difficult
When with conventional programs, the characteristics of ES are:
•
The main conceptual source of ES is knowledge. ES can expand to include a knowledge
acquisition component that processes data and information into rules. Data and information
are indirect conceptual resources.
•
ES does not represent its knowledge in algorithmic-like procedures. AI is based on symbolic
processing of knowledge. Knowledge is represented by symbols such as letters, words or
numbers that represent objects, processes and their relationships.
The knowledge base
contains facts and concepts and relationships among them. Processes are used to manipulate
the symbols in the knowledge base to generate advice or recommendations to solve problems.
•
The objective of ES is to create an efficient knowledge representation to support the inference
mechanism
•
ES keeps the control separate from the rules: the inference engine is responsible for most of
the control. The knowledge base contains the rules or knowledge.
•
Knowledge is expressed as data and not encoded in the control of the ES. This has the
advantage that knowledge can be returned to the user as explanations of the ES conclusions.
•
The order of the rules in the knowledge base does not matter
•
ES provides an explanation facility
•
The rules that comprise the knowledge may be built bit-by-bit, a rule at a time. ES can
tolerate imperfect, incomplete or uncertain business rules or domain knowledge.
•
The ES will produce results even if the knowledge base is incomplete and uncertain. This is
because of the separation of the knowledge and control components.
•
ES frequently uses heuristics in its search approach
•
ES has a reasoning capability, and
•
Maintenance and updating of data are relatively easy. Changes can be made in self-contained
modules.
109
University of Pretoria etd – De Kock, E (2003)
6.1.1 When to use an Expert System
ES is appropriate for a limited number of problem domains. The task should be well bounded and
should not include a large number of event combinations. A general rule of thumb is that if the average
employee can solve the problem in a few minutes it is not worth solving it by an expert system.
ES is appropriate (Raggad & Gargano 1999; Mallach 1994):
•
If great costs are involved when making the decision
•
When applied to repetitive problem domains, the task must be performed often
•
When a big difference exists between the best solution and the worst solution
•
When test data is easy available to test and validate the ES
•
When there is a general agreement on the system’s conclusions. Errors in the input should be
tolerated in the system output. Experts must agree on the solutions.
•
When recognised expertise is available and a necessity during the development of the ES
•
If the problem is clearly specifiable in the area of expertise i.e. what the system is supposed to
do before it is done, the desired system can be developed
•
If the problem is identifiable with a human expert, that is based on human expertise. Experts
must perform the task substantially better than non-experts. The human expertise must be
scarce.
•
If the problem domain is well bounded: for example defining criteria to determine the subject
matter within the system from the matter outside of the system: The task is reasonably stable
and within an acceptable narrow domain.
•
If solving the problem is based on knowledge rather than common sense reasoning: ES is
based on knowledge and naïve systems on reasoning.
•
The solution has to be possible, justifiable and concisely generated (using only a few hundred
rules), and
•
The knowledge has to be of a cognitive style and independent of common sense. Physical
activities are ruled out.
Goodall (1985) provides a comprehensive list of suitable types of tasks appropriate using ES in solving
problems that includes:
•
Interpretation
•
Prediction
•
Diagnosis
•
Debugging
•
Design
•
Planning
•
Monitoring
•
Instruction, and
•
Control
110
University of Pretoria etd – De Kock, E (2003)
Turban et al (2001) calls the above-mentioned types the generic categories of Expert Systems. Most
ES systems will include one of the above problem areas as a dominant characteristic, but may include
some aspects of the others.
6.1.2 Advantages and disadvantages of using ES
Developing an ES is time consuming for the expert, whose knowledge is captured, the knowledge
workers, who will eventually make use of the system; as well as the programmers, who will be
involved in the building of the ES. ES is limited in the sense that it focuses on a narrow problem
domain and can not therefore be applied outside of the specific problem domain. For this reason it
involves intensive development cost.
According to Goodall (1985), Mallach (1994), Turban (1995) and Turban et al (2001) the advantages
of using an expert system (a system with intelligent reasoning) above a conventional information
system (for example a system without intelligent reasoning) are:
•
It improves the quality of the system. The program will function even with incomplete and
uncertain data from the very first time it executes previously non-programmable tasks. This is
an advantage in domains where knowledge changes frequently. It provides consistent advice
whilst reducing the error rate.
More rules can be added to the existing knowledge as
knowledge about the domain increases.
•
It can handle uncertainties expressed as probabilities. The way it is handled depends on the
inference engine of the particular ES.
•
It explains the logic behind its recommendations, making the knowledge explicit
•
It can be used as a training vehicle for users who lack the expertise
•
It provides monetary savings once implemented. ES does not cost money when not being
used. Costs can be cut as the human expert’s availability is not needed all the time.
•
Experts are freed enabling them to focus on tasks requiring their expertise
•
It can codify and preserve knowledge of the specific problem domain. It offers a solution to
scarce expertise where few experts exist for a task or where the expert is about to retire or
leave the job.
•
It increases programmer productivity. The author of the knowledge base will have access to
more advanced techniques and tools than the system’s eventual user. Separating the expert’s
knowledge into modular rules reduces the change of development errors and improves the
maintainability of the system.
•
It can discover new knowledge
•
It can provide increased output and productivity. ES works faster than humans do. Increased
outputs mean fewer workers and reduced costs. ES can be replicated as needed.
•
It eliminates the need for expensive equipment used by human experts for monitoring and
control
•
It makes knowledge and information accessible, and
111
University of Pretoria etd – De Kock, E (2003)
•
It can outperform human experts because of the fact that ES:
o
Make fewer errors
o
Do not become tired or bored and never sleeps.
o
Will not overlook a solution
o
Can handle large volumes of data
o
Can respond more rapidly
o
Can function in hostile environments such as deep-sea drillings and reactor control, and
o
When the knowledge of several experts is integrated, it can outperform any one expert
Another advantage of ES is that it is a tool for manipulating knowledge. It can:
•
Discover new knowledge e.g. discovery of a new prime number
•
Codify and explicitly state old knowledge by forcing the author of the knowledge base to
formalise the expertise in rules or other structures. This resultant knowledge can be used by
anyone who needs it. Tools are available to codify knowledge. One such tool is Teiresias.
•
Assist in teaching once knowledge is codified, and
•
Help analyse knowledge
Mallach (1994), Turban (1995) and Turban et al (2001) give the drawbacks of ES as:
•
The domain of expertise is usually narrow. An ES is designed for a specific purpose and is
not useful for another purpose.
•
ES cannot apply common sense, only its rules
•
Experts do not realise when they reach their limits. They may recommend inappropriate
actions. The experts’ performances “falls off a cliff” rather than degrade gradually.
•
ES may be costly to develop because of the time of human experts and other people involved
in the process
6.2
•
Knowledge is not always readily available
•
Expertise is hard to extract from humans
•
Lack of trust by end-users may be a barrier to ES use
•
Knowledge transfer is subject to a host of perceptual and judgemental biases
•
The work of rare and expensive knowledge engineers is required, and
•
Most experts have independent means of checking whether their conclusions are reasonable
Architecture of ES
6.2.1 The process of ES
Transferring expertise from an expert to the computer and then to the user involves four activities:
knowledge acquisition, knowledge representation, knowledge inference and knowledge transfer
(Turban et al 2001). Knowledge is acquired from experts and other documented sources and then
represented or organised as rules or frames. These rules or frames are electronically stored as a body of
knowledge or a knowledge base. This knowledge base forms one of the distinct parts of an ES. The
112
University of Pretoria etd – De Kock, E (2003)
reasoning mechanism that draws conclusions from the knowledge attained in the knowledge base and
the knowledge obtained from the user forms the other distinct part of the ES.
mechanism is known as the inference engine.
This reasoning
The inference engine results in advice or
recommendations for novices. When understood by the novice, possibly by accessing the explanation
given by the ES, knowledge has been transferred to the user. Figure 2-6 (p21) shows the different
components of an ES and their interaction.
6.2.2 The components of ES
An expert system consists of various major components (Goodall 1985):
•
A dialogue structure
•
A knowledge base, and
•
An inference engine
Turban et al (2001) adds:
•
A blackboard
•
An explanation sub-system, and
•
A knowledge refining system
The user interface provides the conversation of the system with the user. It is responsible for posing
the questions to the user, reading the user’s reply and explaining the rules used to reach a conclusion.
A survey of typical ES suggested that on average (Goodall 1985):
•
Eight percent of the code is used for the inference engine
•
22% of the code is used for setting the knowledge base, and
•
44% of the code is used conversing with the user
The explanation capability is one of the most important features of an ES and it realises in the user
interface. Turban et al (2001) views the explanation facility as a separate component of the ES that can
trace and explain the ES’ behaviour (See Figure 2-6: p21). The body of knowledge (also called the
knowledge base) and the inference engine, the reasoning mechanism of the ES, form the two
fundamental parts of the ES.
These two subsystems provide the ES’ main functionality.
The
knowledge base represents domain specific knowledge in a specific form. The inference engine
deduces facts or draws conclusions from the knowledge base based on the user input and the facts from
the knowledge base and/or other sources. The inference engine and the knowledge base exist as two
separate modules of the ES that work closely together.
♦ The user interface
The knowledge worker (also called the user or decision-maker) gains access to the ES via the dialogue
structure provided by user interface. The knowledge worker provides the ES with input and receives
the system’s explanation of how it reached its conclusion. Output to the user usually comprises of
taking text from the knowledge base and slotting them into a few predefined sentence formats.
113
University of Pretoria etd – De Kock, E (2003)
When allowing the user to use natural language in obtaining input, a misunderstanding could result
because of the ambiguous characteristic of natural language. The use of natural language in the
dialogue with the user relies on the meaning of the user’s utterances. The meaning of what exactly is
meant lies in examining the surrounding context and is prone to be misunderstood. The naïve user is
likely, by the system’s ability to read natural language, to be fooled into thinking that the ES absolutely
understands what he (the user) means. The best way to solve this problem is for the ES to generate all
possible interpretations of any sentence that could be ambiguous, feed it back to the user, and enquire
of the user which interpretation he (the user) really meant. The easiest way to understand the user is
not to allow conversation to be in any natural language, but rather via a predefined syntax, requesting
preformatted specific input.
Questions posed to user and explanations of conclusions that are
reached can be given by taking the text provided by the author of the knowledge base and slotting it
into a few predefined sentence formats.
♦ The knowledge base
The power and effectiveness of the ES is equal to the quality of the knowledge it contains. The
knowledge has to cope with high degrees of complexity and apply the best judgement. The acquiring
of expert knowledge is crucial and involves the gathering of information about a domain usually from
an expert. This information is incorporated in a computer program stored as a knowledge base.
Obtaining knowledge from humans can be a difficult task. Hayes-Roth et al (1983), as referenced by
Turban (1993) and Klein & Methlie (1995), views knowledge acquisition (See Paragraph 6.4: p122)
or knowledge engineering as being composed of five stages:
•
Identification or definition of the problem and the major characteristics of the problem
•
Conceptualisation or the choice of an appropriate representation medium such as rules and the
acquiring of the knowledge associated with the certainty of the knowledge
•
Designing of an architecture or deciding the formalism or acquisition methodology e.g. rules
and the deciding the process extracting knowledge from the experts
•
Implementation, building or the programming of knowledge and the development of a
prototype, and
•
Testing and refining of the knowledge base by subjecting it to examples, examining the
validity of the knowledge.
The methods of knowledge acquisition can be divided into manual, semi-automated and automated.
The primary manual approach is interviewing, ranging from completely unstructured to highly
structured interviews. Automated knowledge acquisition uses an induction system with case histories
and examples as input to derive the knowledge base. It eliminates the role of the knowledge engineer
and expert (Turban 1993). Automatic knowledge acquisition is also known as machine learning.
Knowledge collected must be analysed and coded prior to its representation in the computer.
Automated knowledge acquisition methods are easier to validate and verify.
On average fewer
knowledge mistakes occur when acquiring knowledge from multiple experts because of the enhanced
114
University of Pretoria etd – De Kock, E (2003)
synergy among experts. Difficulties such as meeting time scheduling and resulting compromising
solutions in opinion conflicts may arise.
Knowledge representation is important and crucially affects the ease and speed with which the
inference engine can use it. Knowledge representation implies a systematic means of encoding what an
expert knows about a knowledge domain in an appropriate medium (Goodall 1985). A number of
knowledge acquisition techniques have been developed.
Turban (1993) discussed a variety of
techniques. The selection of a technique depends on the types of knowledge that should be retained in
the knowledge base. Knowledge can be classified as surface knowledge, to put declarative and
procedural knowledge into heuristics to solve a problem quickly; or deep knowledge, which involves
fundamental knowledge of domain including the definition, axioms, general laws, principles and causal
relationships upheld by the knowledge. Surface knowledge is the basis for most common ES using
production rules. Production rules are widely used and quite efficient in diagnosis problems. They are
used to encode rules of thumb also called heuristics or knowledge used by humans (Turban 1995).
Knowledge can be formulated using formal theories or normative models.
Knowledge representation is the systematic means of encoding knowledge of the human expert in an
appropriate medium (Beynon-Davies 1991). Knowledge can be represented as:
•
Predicate calculus or formal logic
•
Business applications in the form of production rules
•
Semantic networks, which organise knowledge through nodes in a graph rather than data
structure and represent relationships between the facts by links between the nodes, and
•
Frames or structured objects that use data structures to store all structural knowledge of a
specific object in one place
Logic itself is not a way for computers to store knowledge, but proves to be a vital tool to think about
how computers store knowledge. Logic is part of mathematics and can be used in various forms to
reason about the correctness of computational representation and inference (Goodall 1985). The forms
of logic include:
•
Programming languages such as PROLOG (programming in Logic)
•
Pro-positional logic or calculus that consist of building blocks such as elementary sentences,
joined by ‘and’, ‘or’ and ‘not’. The internal structure of the elementary sentence is irrelevant.
, and
•
Predicate logic with its basic building block objects and relations such as “is-a” and “has-a”
between them to build statements. The relations is called predicates and deals with the logical
operators “and”, “or” and “not”.
The above provide a theory to formalise the study of reasoning, determining valid knowledge
representations. It is used to prove the correctness or determine that certain types of inference are
correct or incorrect.
115
University of Pretoria etd – De Kock, E (2003)
The forms of knowledge representation often used in ES are:
•
Rules
•
Semantic nets
•
Frames, and
•
Cases
Rules are often called production rules and the program that reasons with rules a production system,
especially when the inference is data-directed forward chaining. Most ES represent knowledge as rules
and therefore the knowledge base is often referred to as the rule base. The reason for its popularity is
that almost every piece of knowledge can be written as a rule. Many ES exist that requires rules as
input and tempts knowledge workers to express knowledge as such. Rules are natural and the only way
to codify some knowledge. Rules are a simulation of the cognitive behaviour of human experts. It
represents knowledge, but also represents a model of actual human behaviour. Rules are easy for a
human expert to read, understand and maintain. If the knowledge is expressed as data and not encoded
in the program’s control mechanism, it can be returned to the user in the form of explanations.
Production rules involve simple syntax that is flexible and easy to understand. They are quite efficient
in diagnosing problems of the form:
•
if (condition)
•
then (conclusion)
Each production rule in a knowledge base implements a chunk of expertise and when fed to the
inference engine, as part of a set, should synergistically yield a better result than any of the rules
individually. In reality, rules are highly independent and adding a new rule may contradict other rules
or cause other rules to be revised. Rule systems can broadly be classified into simple, all rules on the
same level and available to every search cycle, and structured rule-base systems where searches are
limited to segments of the rule base, thus improving the efficiency or the search. A rule set is a named
collection of individual rules pertaining to a distinct aspect of a problem and helps in comprehending
and maintaining the rule base. This structure is a kind of meta-knowledge that is imposed on the
knowledge base. Each sub problem could have its own rule set (Klein & Methlie 1995).
Semantic Nets is a popular and easy-to-understand way of representing non-rule knowledge. Semantic
networks organise knowledge through nodes in a graph rather than a data structure. Links or arcs
present relationships between the named nodes. The links or arcs represent relationships such as is-a,
has, is, own, needs and reflects the association between concepts. An ES that stores knowledge as a
semantic net incorporates an inference engine devoted to operations like inheritance. Such an inference
engine will consist of two parts. One part will deal with rules by forward chaining, backward chaining
or some other method. The other part will handle net operations matching relevant links in the net to
deduce facts.
116
University of Pretoria etd – De Kock, E (2003)
Objects can be described by a number of features or attributes, many of which stay constant from one
instance to the next. A frame is a piece of structured data about typical characteristics of an object, act
or event. The knowledge is more descriptive than procedural. Similar to rules, frames must be able to
deal with uncertainty and missing values. Frames may have default values and slot-filling procedures
associated with the slots, to cope with missing values. Frames enable reasoning about object features
such as inheritance and the occurrence of exceptions. The reasoning process starts by identifying a
frame as applicable to the current situation. Matching the set of frames against the facts available
selects an appropriate frame. The use of frames is relevant to non-monotonic logic. Non-monotonic
reasoning expresses reasoning with default attributes.
Case-based reasoning is a process that uses similar problems to solve the current problem. It consists
of two steps:
•
Find those cases in memory that solved problems similar to the current problem, and
•
Adapt previous solutions to fit the current problem
The case library forms an extra important component in case-based reasoning. The inference cycle
using CBR consists of:
•
Retrieving solutions
•
Adapting solutions, and
•
Testing solutions
The critical step is to find and retrieve a relevant case from the case library. Cases are stored using
indexes. The stored case contains a solution, which is then adapted by modifying the parameters of the
old problem to suit the new situation resulting in a proposed solution. The solution is tested and if
found successful, added to the case library (Klein & Methlie 1995). Knowledge acquisition is easier in
case-based reasoning because of the granularity of the knowledge.
Knowledge is presented in
precedent or resultant cases.
Beyond the knowledge representation language (rules, semantic nets, frames, cases), the knowledge
engineer needs further aids such as tools to edit the knowledge base; inference tracers to assist in error
detection; and analytical tools to find, update and consistently check the represented knowledge or
attributes (Klein & Methlie 1995).
♦ The inference engine
The mechanism that performs the search and reasoning in rule-based systems is called the inference
engine. The inference engine is activated when the user initiates the consultation session. The
inference engine finds the rules that match the given facts, selects which rule to execute and executes
the rule by adding the deduced fact to the Working Memory (WM). The inference engine uses pattern
matching to select the qualifying rules. The choice of which rule to fire is done by conflict resolution.
The most commonly used conflict resolution strategy is the first found strategy. The first applicable
117
University of Pretoria etd – De Kock, E (2003)
rule is executed (Klein & Methlie 1995) or fired by applying rule deduction or using formal logic. A
new fact is concluded and added to the working memory and new patterns found that match the new
fact. This sequence of steps and the linking of facts and patterns and rules are known as chaining
(Klein & Methlie 1995).
The inference engine deduces facts or draws conclusions from the knowledge base based on the user
input and the facts from the knowledge base and/or other sources. Three basic techniques are identified
when inferring facts or conclusions from the knowledge base:
•
Forward chaining
•
Backward chaining, and
•
Hybrid chaining: using both forward and backward chaining
New knowledge can be inferred from the existing knowledge in the knowledge base. The specification
of how the reasoning is done is the core of the ES. This programming task is independent of the
subject knowledge in the knowledge base component. The reasoning method treats facts as arbitrary
symbols. The same inferencing method can be used by different kinds of knowledge (Goodall 1985).
Various algorithms can accomplish this task. One such algorithm used by ES, is the Rete Algorithm
(See Paragraph 4.2.9: p45)
Forward chaining match the current state with the rules’ antecedents or conditions in the knowledge
base. If the condition is true, the inference engine adds the conclusion to the list of known facts.
Known facts imply other facts. This technique is also known as data driven. It can be very inefficient,
especially if many rule conditions match the data provided by the user. The data provided by the user
is examined and new knowledge is concluded or asserted and responded to.
The second technique, backward chaining, also known as goal driven or directed, is more efficient if
the number of goals is limited and involves the acquiring of information from the user in the form of
questions to draw specific conclusions. This style is used to confirm diagnostic tasks and works
backwards from what needs to be proved to the facts that might affect it. It is a goal-directed backward
chaining of rules to try to prove a hypothesis made by the system. Sometimes stating the conclusions
and trying to prove them are unhelpful because none of the multiple goals exist. If this is the case,
forward chaining should rather be considered
When a large problem domain is involved, a more efficient program is yielded when the two
techniques are used in combination - hybrid chaining. To model the strategies used by a human
expert, an ES has to employ complex inference that almost always include both forward and backward
chaining with potentially many variations. One inference engine will not suit all possible tasks solved
by an ES. A typical rule-based system represents heuristic knowledge: that is short cuts in the
reasoning procedure. This process uses shallow knowledge.
118
University of Pretoria etd – De Kock, E (2003)
Model-based reasoning is based on fundamental or deep knowledge of the problem domain. The
relationships modelled could be either structural or behavioural. Structural relationship models are
used for advice systems and behavioural relationship models for Decision Support Systems that do
scenario analysis such as a “what-if” analysis. The behavioural model consists of causal relationships
that can be quantitative or qualitative (Klein & Methlie 1995).
In model-based reasoning, the
reasoning follows the fundamental relationships, the cause-and-effect paths using a deeper knowledge.
The reasoning styles are directed by what need to be proved (goal directed) versus by what data are
available (data directed). When pursuing the goal directed style and several rules exist that prove some
conclusion the rule with the best change of proving it, is selected. This is done by conflict resolution,
for example a more specific rule is chosen above a less specific rule.
To simplify the process each rule can be assigned a number to reflect how useful it is. The number
could be a guide to select one rule from several that proves the same conclusion. Rules that deal with
uncertainty or imperfect data or are referred to as approximate rules, are called probabilistic
reasoning. An integral part of the working of the ES is how the inference engine determines when to
enquire the appropriate information from the user. Probabilistic reasoning is a method used by an ES
to draw inferences from the domain and problem knowledge where the knowledge or its
implementations or both are less than certain.
Three ways exist to represent uncertain knowledge:
•
A single subjective probability is assigned to the proposition, possibly derived using Bayesian
inference. Bayes’ theorem provides a way of computing the probability of a particular event
given some set of observations. The main fact here is not the derivation of the value, but the
association of the specific proposition with a single value. Criticism against this approach
include that the single value will not reveal much about the rules precision and yet another
states that this single value combines the evidence for and against this proposition without
reflecting how much of each there is. An alternative approach is to provide an independent
measure of belief and an independent measure of unbelief that can be combined into a single
certainty factor.
•
Fuzzy logic is another technique that can handle inexact data, and
•
A final method is the use of a few quantitative qualifiers to express the levels of probability.
Possible values may include unlikely, possible, likely, impossible or certain
The knowledge base can include meta-rules. A meta-rule is a rule about the rules and potentially
affects the sequence of all other rules called. Meta-rules make the knowledge base more complex and
more difficult to understand e.g. the use of different rule sets. Two types of facts exist: those received
from the user in the form of input data (known facts) and those deduced from the input data (inferred
from primitive data). Knowledge can also be obtained from other sources such as databases (DB).
Much research is connected to Expert Systems and databases to refrain from transcribing some of the
information into rules.
119
University of Pretoria etd – De Kock, E (2003)
In acquiring information from databases, difficulties may be experienced such as:
•
Connected DB may not be able to provide answers to all the queries ES would like to put to it.
The ES needs to distinguish between questions viable to put to a DB and that not.
•
The ES and the DB may not be able to run in the same extensive memory needed for an ES
and a DBMS on one machine, and
•
Queries to the DB need to be phrased in a special DB query language. The ES should be able
to formulate the query and interpret the reply.
The rules connect items of knowledge about a problem. Some ES concentrate more on facts and less
on rules. Many pieces of information may have more than one attribute. Such knowledge is said to be
structured, compound or multi-attributed. All ES languages should provide for it in some way.
ƒ
Inferencing using production rules associated with objects
The following describes the interaction between the knowledge base and the inference engine:
A production system typically consists of a rule base (knowledge base) and a rule interpreter (inference
engine) that decides when to apply the rules to the data, goals and intermediate results in the working
memory. The Working Memory (WM) holds the Object-Attribute Values (OAV) used to drive or fire
the rules. The OAV triggers some rules in the WM by satisfying their conditions.
ƒ
Inferencing using frames
Non-monotonic reasoning expresses reasoning with default attributes designed for frames.
The
process retracts rules when proven wrong. Non-monotonic reasoning is a mathematical tool by which
default information holds until this information becomes inapplicable.
♦ Blackboard subsystem
Blackboards are not an alternative to frames and nets for storing knowledge. It is rather a complex
method used in complex systems to record choices, guesses and decisions about what to do next. It is
used in situations where more than one source of information is active. All knowledge sources have
access to a shared database. It involves an architecture that allows the independent knowledge sources
to communicate through the central device: the blackboard.
When a knowledge source is activated, it examines the current state of the blackboard and implies its
knowledge to either create a new hypothesis or change an existing one. The scheduler determines
which knowledge source can contribute to the solution in the blackboard next. Hearsay-III - a speechunderstanding project; ExperTax - a tax planning system, and FINSIM are examples of blackboard
systems (Klein & Methlie 1995). Turban et al (2001) describes the blackboard as a kind of database
that an area of working memory set aside for the description of the current problem, the input data and
the intermediate results.
120
University of Pretoria etd – De Kock, E (2003)
♦ An explanation sub-system
A by-product of the inference engine is that it is able to provide a trace of rules (subject knowledge)
used and thus provide an explanation of its conclusions. This explanation can be presented to the ES
user in the dialogue component. The explanation sub-system provides an area where the behaviour of
the ES can be accounted for (Turban et al 2001). It provides answers to questions such as:
•
How was a certain conclusion reached?
•
Why was a certain question asked?
•
What is the plan to reach the solution? , and
•
Why was a certain alternative rejected?
♦ A knowledge refining system
Human experts can analyse their own performance, learn from it and improve it for future
consultations. Having the program evaluate the reasons for success or failure and learning from it
would greatly enhance ES. This functionality is not available in many commercial Expert Systems yet,
but is being developed in experimental ES (Turban et al 2001).
6.2.3 The process of ES
The process of ES (See Figure 2-6: p21) can be divided into two parts (Turban et al 2001): the system
development process in which the ES is constructed and the consultation process which describes
how advice is rendered to the users. The development process starts with the knowledge engineer
acquiring the knowledge from the expert.
This acquired knowledge is then programmed in the
knowledge base as facts about the subject and knowledge relationships in terms of if-then rules.
The consultation process involves the user who starts the process by requiring advice from the ES.
The ES provides a conclusion and explanation, using its inference engine. The engine searches the
knowledge base for an appropriate action by matching the input the user provides to the facts and rules
in the knowledge base. To execute this task the inference engine uses a work area or temporary
databases called the blackboard. All intermediate results are stored here. An explanation of actions
taken by the inference engine can be provided to the user. Finally, the knowledge in the knowledge
base may be refined by repetitive consultations. This knowledge refining system is not available in
many Expert Systems now, but is being developed as experimental ES.
6.3
ES and expert shells
An expert shell is an ES without the domain-specific knowledge (Beynon-Davies 1991). Mallach
(1994) refers to a shell as a pre-packaged inference engine. Construction time can be reduced by using
an ES shell. Advantages of using a shell include (Beynon-Davies 1991):
•
An expert shell can assist in rapid prototyping. The programmer needs only construct the
knowledge component. The structure of the ES exists, enabling the developers to concentrate
121
University of Pretoria etd – De Kock, E (2003)
on the substantive content rather than the form of the ES. A shell helps to reduce the level of
skill required by the developers by effectively supplying some of the required expertise.
•
A shell imposes prior structure, enabling developers to concentrate on substantive content
rather than form, and
•
6.4
A shell reduces the required skill level of Expert Systems builders
Knowledge acquisition and knowledge base construction
The knowledge that is contained in the system determines the effectiveness of the ES (See Paragraph
2.3.4: p22). Knowledge engineering may contain the following steps (See Paragraph 6.2.2: p113):
•
Definition of the problem
•
Acquisition of expert knowledge
•
Design of an architecture, and
•
Testing and refining of the program
Acquiring expert knowledge is a crucial component of knowledge engineering. This phase is difficult
and time consuming. It is the process of gathering the relevant information about a domain, usually
from an expert. Usually a computer program is compiled to incorporate all the knowledge of the
domain. A number of knowledge acquisition techniques have been developed. As mentioned before
knowledge can be classified as surface knowledge or deep knowledge. Surface knowledge involves the
presentation of declarative and procedural knowledge into heuristics to quickly solve the problem.
Surface knowledge is the basis for most ES and uses production rules. Deep knowledge involves
fundamental knowledge of the domain including definitions, axioms, general laws, principles and
causal relationships of the domain.
The human expert, from whom the knowledge is acquired, needs to be a person that can effectively
cope with the problem domain. Selection of the person is based on reputation. The expert should be
available and willing to co-operate in the process of developing the ES especially when acquiring the
knowledge and testing the system. The expert must be able to communicate his expertise.
Validity of knowledge in any system is crucial. If the knowledge inferred proves to be faulty or
erroneous in any way, the wrong decisions made by the experts will be made more speedily than when
the process was a manual one. The sources of errors could include:
•
Not much is known about the problem domain; too little knowledge is available. To solve the
situation more domain expertise need to be obtained. , and/or
•
The environment of the decision might be changing in which case the captured knowledge
needs to be maintained
Relational database systems manage knowledge in the form of facts but are quite different when
compared to knowledge bases. To understand ES and knowledge base construction, Beynon-Davies
(1991) compares the two technologies.
122
University of Pretoria etd – De Kock, E (2003)
6.4.1 Comparing a Knowledge-based System (or an ES) to a Relational
Database System
Beynon-Davies (1991) states that knowledge can be represented as both facts and rules. Facts declare
relationships between objects. Rules define the process by which new facts are generated from old
facts. Rules are mechanisms to manage the information explosion in the attempt to represent reality.
Beynon-Davies (1991) declares a database as a collection of structured data shareable between
different parts of an organisation’s information system, having properties such as program-data
independence, data integration, data integrity and separate logical and physical views of the same data
using a specific underlying data model. Hoffer et al (2002) calls these different models the database
architecture. The different styles of database management characterised by the way data are defined
and structured are called the database architecture. A database management system supports one of
five architectures: Hierarchical, network, relational, object-orientated or multi-dimensional database
models.
Beynon-Davies (1991) compares knowledge-based systems to the relational model of
databases when discussing the integration of expert and database technology. The relational model
consists of tables representing the information stored on objects or entities and the relationships
between them.
Important characteristics of relational databases are:
•
Data are stored in a component called a database
•
To use one data structure called a table or relation
•
To handle large collections of facts representing declarative knowledge or relationships
between objects
•
Rules can be represented in databases only in a declarative manner, storing values in a
separate table reporting the relationship. This may result in many tables to represent all the
different relationships that exist between objects.
To limit the number of tables some
knowledge is embedded in the high-level language programs that accesses the database called
the front-end.
•
Rules as procedural knowledge are represented as constraints or additional tables mainly to
govern data integrity
•
A high-level language code interacting with the database contains and processes the
procedural knowledge
•
A database is a limited Knowledge-based System. It is a subset of what is normally meant by
knowledge.
•
A RDBMS (Relational Database Management System) is program-data independent. Data
can be maintained independently of the programs that use it.
•
Addition of a rule involves modifying the structure of the database (addition of a table
structure) – not a simple matter
•
No inferencing except when explicitly programmed
•
Data must be valid. Rules govern the logical consistence of the database.
123
University of Pretoria etd – De Kock, E (2003)
•
A database uses a tool called the database management system (DMBS) to build database
systems
•
The DBMS (Database Management System) is built upon a data model or architecture, and
•
The database management system (DBMS) is an integral part of the database and handles
manipulation of large collections of facts
In contrast with DBMS, the characteristics of ES are:
•
Knowledge (facts and rules) are kept as a separate component called a knowledge base
•
The data structure depends on the knowledge representation chosen e.g. production rules,
structured objects or frames and predicate calculus and others
•
ES handle large collections of rules representing procedural or derivation of new facts from
existing facts
•
Rules as declarative knowledge are added independently from other rules or the process that
uses it
•
Rules as procedural knowledge are stored as data mainly used by the process to derive new
facts from existing facts
•
Knowledge can be represented as both facts and rules
•
The knowledge base’s general-purpose processor activates the procedural knowledge
•
ES are knowledge-process independent. Rules are maintained independently of the process
itself.
•
Addition of a rule involves adding or deleting a condition – a relatively simple matter
•
ES separate inference and knowledge. Once the inference process is in place, the knowledge
worker can focus on the knowledge itself. ES is sometimes referred to as a knowledge-based
system because knowledge is kept in a separate component as data.
•
Knowledge may be incomplete
•
ES use a tool called an expert shell to build Expert Systems
•
An expert shell is built on some formalism for knowledge presentation, and
•
The inference mechanism as an integral part of ES handles the activation of large collections
of rules
A database system is an example of a limited Knowledge-based System (Beynon-Davies 1991). A
knowledge base handles large amount of rules (procedural and declarative knowledge), while a
database handles large amounts of facts or declarative knowledge. The procedural knowledge of a
database is found in the associated external high-level language interacting with the database. The
procedural knowledge (rules) are stored as data in a Knowledge-based System and activated by an
inference process integral to the Knowledge-based System.
6.5
Summary
Expert Systems are based on two fundamental principles: the appropriate representation of domain
knowledge and the control of the domain knowledge. A standalone expert system stores a few facts
124
University of Pretoria etd – De Kock, E (2003)
and relies on the user to pass information to the system when needed to perform certain deductions.
The combination of Expert Systems and database technology could benefit both technologies. Models
have been developed for data base systems that would benefit the storage of knowledge in a knowledge
base. DBMS technology could contribute Expert Systems by giving them the ability to access large
collections of facts and also apply features such as concurrency control, data security and optimised
access to knowledge base items (Beynon-Davies 1991).
A deductive component could enhance database technology by providing the rules component of an ES
(Beynon-Davies 1991). A frame is roughly the equivalent to a row in a relational database. A slot is
roughly the notion of an attribute in a relation. Frames and slots are artefacts used in ES rules to
represent knowledge. A frame (See Paragraph 6.2.2: p113) use slots and is a knowledge presentation
used in Knowledge-based Systems.
The writer concludes in agreeing with Turban (1995) that an ES component is ideal to assist a decisionmaker in an area where expertise is required.
125